An AI memory gap appeared after just one week: people who had developed ideas with a chatbot often remembered the ideas themselves, yet struggled to recall whether the human or the AI had created them. In a preregistered experiment with 184 adults, source accuracy for ideas fell to 37.7 per cent when AI supplied the idea and the participant wrote the elaboration. For ideas created without AI, it was 92.4 per cent.
The finding is more consequential than ordinary forgetfulness. AI tools are entering work precisely where authorship, responsibility and learning matter: brainstorming, research, planning, writing and design. If users retain a useful idea but lose track of its origin, they may overstate their own contribution, fail to recognise outside influence or become unable to explain how a decision was reached.
How the one-week memory test worked
The researchers recruited 184 English-speaking adults through Prolific, including 112 people in the UK and 72 in the US. Most reported using AI frequently. In the first session, each participant worked through eight problem statements and produced 40 short ideas with one-sentence elaborations. The experiment alternated between work completed without AI and work involving a GPT-4.1-mini chatbot.
The design separated two parts of creative work that are often bundled together. Sometimes the person generated an idea and AI elaborated it. Sometimes AI generated the idea and the person elaborated it. In other trials, the person or the AI produced both. That gave the researchers four authorship combinations and made it possible to test whether people remembered the source of an idea differently from the source of its supporting explanation.
About a week later, participants saw their 40 earlier items mixed with 20 plausible distractors. For each item, they were asked whether they had seen it before, who had created the idea, who had written the elaboration and how confident they were. The delayed test matters because most real accountability questions arise after the drafting session, not while a model’s reply is still visible on screen.
People remembered the work better than its source
Recognition of the material was generally strong, but source memory changed sharply with the division of labour. When people created both the idea and its elaboration, they identified the idea’s source correctly 92.4 per cent of the time. When AI created both parts, idea-source accuracy was 79.3 per cent. The hardest condition was mixed authorship, especially when AI supplied the idea and the human developed it.
In that AI-idea, human-elaboration condition, correct source attribution for the idea fell to 37.7 per cent. When the human supplied the idea and AI elaborated it, idea-source accuracy was 64 per cent. Attribution for elaborations was better, ranging from 79 per cent to 91.5 per cent across conditions, but it still weakened when responsibility was split.
Across the experiment, participants correctly identified the source of ideas 65.8 per cent of the time and the source of elaborations 72.8 per cent of the time. Their confidence did not fully track that weakness. The researchers found that participants overestimated idea-source accuracy by 12 percentage points and elaboration-source accuracy by six points. People were not merely uncertain. They were more certain than their performance justified.
The study’s University of Bayreuth summary describes the effect as a problem of cognitive provenance: remembering not only information, but where it came from and who contributed it. That is the mental record people use when they decide who deserves credit, which sources need checking and whether a conclusion rests on their own reasoning.
Why mixed authorship is especially slippery
A finished idea can feel personally owned after someone has selected it, rewritten it and connected it to a wider argument. The experiment cannot prove the precise psychological mechanism, but its pattern is consistent with that kind of integration. Human elaboration may make an AI-originated idea feel more self-generated because the participant has invested effort in developing it.
The reverse arrangement creates a different ambiguity. A person may originate an idea, then encounter a polished AI elaboration that becomes the most memorable wording. Later, the fluency of the completed item can obscure which part belonged to whom. The result is not simple theft or deliberate misrepresentation. It is a memory problem created by a workflow in which authorship changes from sentence to sentence.
This distinction complements LiveAIWire’s earlier report that AI assistance can hide weaker human skills. In both cases, the visible output is a poor measure of what the person contributed. A strong answer may conceal limited unaided ability, while a remembered idea may conceal uncertainty about whether the person originated it.
What the AI memory gap means for you
For an individual user, the most useful safeguard is to preserve provenance while the work is happening. Keep AI-generated suggestions visibly distinct from your own notes, save important prompts and outputs, and record the reason you accepted or rejected a suggestion. A colour, comment or short source label can do more for later accuracy than relying on memory after a document has been repeatedly edited.
Writers and researchers should treat AI provenance like ordinary sourcing. If a model introduces a claim, quotation, legal authority or study, locate the original source before using it. If it contributes only an angle or metaphor, note that contribution where your professional or academic rules require disclosure. The goal is not to assign every connecting word. It is to preserve a defensible account of material influence.
Teams need shared rules because individual note-taking breaks down when drafts circulate. Version history, tracked changes and labelled AI passages make it easier to reconstruct responsibility after an error. They also protect human contributors from having their ideas casually attributed to the tool. Provenance is not only a constraint on AI use. It is a record of human work that might otherwise disappear.
Authorship affects accountability, not just credit
Misremembering an idea’s source can change how confidently it is used. People generally apply more scrutiny to an outside suggestion than to a conclusion they believe they reached themselves. If an AI-originated idea is later remembered as self-generated, the mental cue to verify it may vanish even though the model could have produced it from weak or false information.
That risk becomes serious in medicine, law, finance and public administration, where a reviewer may need to reconstruct who made a recommendation. A workflow can formally leave a human in charge while making genuine responsibility hard to locate. The person signs the document, but may no longer know which premises came from the system or which alternatives the system quietly excluded.
The same provenance problem can narrow creativity. LiveAIWire’s analysis of AI-assisted ideas becoming more similar showed that an individual benefit can coexist with collective convergence. If users also forget which themes the model introduced, they may mistake shared machine influence for independent agreement and underestimate how much a common tool has shaped the group.
This is not the same as an AI-induced false memory
The new study concerns source attribution, not the creation of an event that never happened. Participants often recognised real material but confused whether it came from themselves or the chatbot. By contrast, a separate experiment covered by LiveAIWire found that a misleading AI interviewer increased false recollections about a robbery video. The two risks operate differently.
One changes the content of memory; the other blurs its origin. They can nevertheless reinforce each other. If a chatbot introduces a false detail and a user later remembers that detail as their own inference, correction becomes harder. The user may defend it as personal recall rather than recognising it as an external suggestion that should be checked.
LiveAIWire’s broader examination of AI and human memory considered how tools can support recall while also changing what people practise remembering. The latest experiment adds a practical warning: an external memory aid can preserve content yet weaken the map showing where that content originated.
The study has important limits
The experiment tested one constrained ideation task, one chatbot interface and one fixed delay. Participants worked in a research setting without the folders, chat histories, colleagues and audit tools available in many workplaces. The results therefore do not establish how often source confusion occurs in a live newsroom, design studio, classroom or board meeting.
The model was GPT-4.1-mini, and different interfaces could create different memories. A tool that inserts text directly into a document may blur authorship more than a separate chat window, while an interface that automatically labels generated passages may preserve it. Longer projects could deepen the confusion through repeated editing, or improve accuracy because contributors leave more records.
The peer-reviewed CHI 2026 record supports the study’s publication status, while the public preprint provides methods, results and revision history. The evidence is strong enough to identify a source-memory problem under the tested conditions, but not to claim that every AI user routinely forgets authorship.
AI systems need provenance people can actually use
Technical logs alone will not solve the problem if they are too difficult to inspect. Useful provenance should travel with the work: visible labels for generated text, accessible version histories, links back to the relevant exchange and a clear record of human edits. The record must help someone answer a practical question months later, not merely satisfy a compliance box.
Good design can also introduce a brief act of reflection. A system might ask users to mark which ideas they are adopting, summarise why they chose them or review a contribution map before exporting a final document. Those small steps preserve active judgement and create memory cues without forcing people to archive every token of a conversation.
The central lesson is not that AI makes authorship impossible. It is that intuitive memory is a weak audit trail for collaborative work with machines. One week was enough for many participants to remember the answer more clearly than the author. Organisations that care about originality, learning or accountability should build the record at the moment of creation, before a fluent shared draft makes every contribution feel like one voice.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.